Multi-Scale Attention Network for Building Extraction from High-Resolution Remote Sensing Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Building extraction from high-resolution remote sensing images is crucial for urban planning, resource management, and environmental conservation.
• Deep Neural Networks (DNNs) show promise but often neglect spatial information and suffer from noise amplification during feature fusion.

Purpose of the Study:

• To develop an advanced deep learning model, the Multi-Scale Attention Network (MSANet), to overcome limitations in current DNN-based building extraction methods.
• To improve the accuracy and robustness of building extraction by effectively utilizing multi-scale spatial and channel information.

Main Methods:

• Implemented a Multi-Scale Attention Network (MSANet) incorporating multi-scale channel and spatial attention mechanisms for feature extraction.
• Utilized adaptive hierarchical weighting and a gating mechanism for effective multi-scale feature fusion.
• Evaluated MSANet on the WHU aerial and WHU satellite image datasets.

Main Results:

• MSANet achieved high performance metrics on both datasets.
• Achieved an F1 score of 93.76% and an IoU of 88.25% on the WHU aerial imagery dataset.
• Achieved an F1 score of 77.64% and an IoU of 63.46% on the WHU satellite dataset II, outperforming DeepLabV3 and GSMC.

Conclusions:

• The proposed MSANet effectively addresses the limitations of existing DNN models for building extraction.
• MSANet demonstrates superior performance in extracting building features from high-resolution remote sensing images, offering a valuable tool for various applications.